German grammar taught as a type system. A static React app with a focused, chapter-by-chapter Learning Mode and an unrestricted Free Mode: a 32-module A2–C1 reference, a drill engine that generates questions fresh and weights them toward your weak spots, an adaptive CEFR placement exam, and a writing trainer that sends your text to Claude for a graded, rule-by-rule correction.
No backend. Everything runs in the browser and progress stays in localStorage.
| Übersicht | Regeln | Drill |
|---|---|---|
 |
 |
 |
| Einstufung | Schreiben | Einstellungen |
|---|---|---|
 |
 |
 |
npm install
npm run devBuild a static bundle:
npm run buildThe output lands in dist/. vite.config.js sets base: "./", so the same
build works at username.github.io, at username.github.io/repo/, and from the
local filesystem — no path configuration needed.
Learning Mode is enabled by default. To build the original Free Mode only and hide all learning-path controls:
VITE_LEARNING_PATH_ENABLED=false npm run build.github/workflows/deploy.yml builds and publishes on every push to main.
One-time setup: in the repository, go to Settings → Pages → Build and
deployment → Source and choose GitHub Actions. Then:
git init && git add -A && git commit -m "Deutsch A2 → C1"
git branch -M main
git remote add origin git@github.com:<user>/<repo>.git
git push -u origin main| Bereich | Was es tut |
|---|---|
| §0 Lernpfad | One guided learning window from A1/A2 foundations to C1. It sequences explanation, preconfigured exercises, and a short writing application without sending the learner to other tools. |
| §1 Regeln | 32 modules A2→C1. Each one: the rule as code, the table it collapses to, real examples, and the mistakes that cost marks in an exam. |
| §2 Drill | 21 generators produce questions on demand. Every answer — right or wrong — prints the derivation that produced it. Weighted toward your weakest rules. |
| §3 Einstufung | 30 adaptive questions. Climbs after three correct at a level, drops after two wrong. Returns a CEFR estimate, the gap to C1, and your weakest topics. |
| §5 Spickzettel | 24 exam task types taken from the Goethe, telc, TestDaF and DSH model papers — each with the mechanical trick that solves it, the traps built into it, a wrong-next-to-right example, and what changes about it at A2, B1, B2 and C1. Plus part-by-part blueprints of eight real exams and a per-level page of what an examiner expects. Available in both modes. |
| §4 Schreiben | 12 exam-style writing tasks A2→C1. Claude returns a CEFR estimate, marks per criterion, every error with its rule, correct-but-below-level phrasings with their target-level version, and your text rewritten at level. |
| §5 Einstellungen | API key, mode, export/import, reset. |
src/engine/grammar.js derives forms rather than storing them: articles and
adjective endings from the case/gender tables, all six tenses plus Konjunktiv I
and II and the full passive paradigm by composition. Konjunktiv I applies the
real fallback rule — if the form would be identical to the indicative, it
switches to Konjunktiv II and says so.
src/engine/drills.js turns that into questions. src/engine/exam.js runs the
adaptive ladder and estimates the level.
Ask questions to the teacher (powered by Claude) and learn more how to write and understand German.
Frau Müller is always there to give you a hand and answer your question!
"Code" in the German language, like you would be learning a programming language, get highlighted chars which will enable you to understand the syntax. Let your imagination flow and get corrected by AI giving you recommendations on how to improve your german writting and what are you fails.
Get recommendations to improve your text:
Practice based on your own mistakes, and learn how to do it right and why.
Two modes, switchable in §5 Einstellungen.
The app builds the full prompt including the JSON schema. You paste it into claude.ai, the Claude app, or Claude Code, and paste the JSON answer back. The app parses and renders it exactly as if it had called the API. Nothing leaves the page except what you copy yourself.
This is a static site with no server, so the browser calls the Anthropic API directly. That means:
- The key is stored in this browser's
localStorageand is visible in the network tab on every request. - Anyone with access to the browser profile can read it.
- Use a key scoped to you, don't use this mode on a shared machine, and rotate the key if you ever publish the site somewhere others use it.
- A progress export never contains the key — it is stored under a separate
localStorageentry and excluded from the export. - The header and API settings count input, output, and cached tokens used by this app for each key. The available figure is Anthropic's latest rolling token rate-limit value; standard API keys do not expose an account credit balance. The local counter can be reset independently of learning progress.
The API path uses the official @anthropic-ai/sdk with
dangerouslyAllowBrowser: true, claude-sonnet-5, and structured outputs, so
the response is schema-validated JSON rather than prose to parse.
Get a key at console.anthropic.com. Requests are billed to your own account.
- Correct writing — CEFR estimate with reasons, 0–5 per criterion
(Aufgabe, Kohärenz, Wortschatz, Grammatik, Register), every error as
original → correctedwith the rule behind it and a severity, up to three below-level phrasings with their target-level version, and the whole text rewritten at level. The correction schema is scoped to the selected CEFR level: higher-level grammar cannot be returned as an assessed error or used to block course progression. - Generate challenges — 8 fresh questions aimed at your weakest rules. The built-in generators have fixed patterns; Claude does not.
- Explain a mistake — on a wrong drill answer: the rule, why your answer fails, a minimal pair, and a mnemonic.
Learning Mode deliberately shows one learning window plus Settings. Related chapters are grouped into 11 named learning blocks such as A2.1 and B1.2. Each chapter contains an introduction, explanation, and preconfigured exercises; a rolling checkpoint requires 3 correct answers out of the latest 4. The learner then continues with the next connected rule. A short writing application appears only at the end of a block (usually after 2–3 chapters) and deliberately combines all rules in that block. This avoids gating progress on a text before its supporting grammar has been taught. Before the guided path advances from A2 to B1, B1 to B2, or B2 to C1, a mandatory level exam asks one question from every chapter in the completed level and requires an 80% score. Failed attempts identify the chapters to review and remain retryable; passing attempts are saved with the learning progress. The header reports the current block, CEFR level, and total path progress. In Settings, the learner can jump directly to A1/A2, B1, B2, or C1 if browser progress was lost; jumping does not falsely mark earlier chapters complete.
Free Mode is the original unrestricted interface. Every reference, drill, exam, and writing tool remains available. Users can switch modes in Settings without losing either kind of progress.
Every Learning Mode chapter includes an optional, clearly labelled KI-Lernbegleiter. It builds a compact anonymous profile from rule accuracy, recent checkpoint results, completed chapters, the last placement level, and error-category metadata from prior corrections. Raw learner writing is not sent when creating this profile.
The model returns a structured personalization pack for the current chapter: an individual focus with evidence, an adapted explanation and memory hook, targeted examples, four warm-up exercises, and a tailored block application that is shown when the learner reaches the end of that learning block. The pack is cached in browser progress, so it does not need to be regenerated on every visit. API-key and manual copy/paste modes are both supported.
AI output cannot change the CEFR sequence or pass a learner. AI exercises are explicitly treated as warm-up material; only the built-in grammar generator and its fixed 3-of-4 checkpoint update mastery and unlock progression. The complete built-in lesson remains available when AI is disabled or unavailable.
Stored in localStorage under dc1:progress: learning-path chapter/checkpoint
state, cached AI personalization packs, per-rule mastery, streaks, daily counts,
exam history, and every submitted text with its correction.
- Backup / Restore — a versioned JSON file in §9 Settings. It contains the complete learning state: settings, learning path, answers, mastery, streaks, exams, submitted texts and corrections, mistake-practice attempts, cached AI exercises, and teacher-chat history. Restoring shows a summary and requires confirmation before replacing the browser's current state. Older exported progress files remain importable. The API key is always excluded.
- Reset — §9, behind a confirmation. Clears mastery, streaks, exams and texts. The API key is separate and is deleted with its own button.
src/
data/
lexicon.js nouns, verbs, modals, prepositions, connectors, phrase banks
curriculum.js 32 grammar modules A2→C1
writing.js 12 writing tasks + the CEFR marking criteria
engine/
grammar.js declension + conjugation, each with its rule trace
drills.js 21 question generators + weak-rule weighting
exam.js adaptive ladder, fixed item bank, CEFR estimator
lib/
claude.js Anthropic SDK calls + the manual copy/paste prompts
storage.js localStorage, export/import, reset
components/ Dashboard · Learn · Drill · Exam · Write · Settings
If this project helps you learn German, you can buy me a coffee with PayPal.
Made with ♥ by Mathias Brunkow Moser.
GPL-3.0-or-later — see LICENSE.
The code and content in this repository were generated with AI assistance (Claude Code, Anthropic), directed and edited by the author.
AI makes mistakes, most of the content was verified by a German Teacher